SkillFlow：用流匹配信号驱动 Agent 技能自主进化，14 个基准全面超越

Agent 技能库该怎么管理？靠 LLM 自己判断”何时创建、何时剪枝”，往往退化成拍脑袋。SkillFlow 给出了一个基于流匹配训练信号的数学答案——在 14 个基准上全面超越现有方法，EM 平均提升 41.2%，计算成本反而降低了 32-35%。

📌 论文：SkillFlow: Flow-Driven Recursive Skill Evolution for Agentic Orchestration
arXiv:2605.14089 全文地址：https://arxiv.org/abs/2605.14089
代码仓库：https://github.com/beita6969/SkillFlow
项目主页：https://skill-flow.org

现有方法的三个根本问题

在 SkillFlow 之前，LLM Agent 的技能演化面临三个未被解决的系统性缺陷：

1. 奖励最大化下的策略坍缩（Strategy Collapse）
传统 RL 训练会收敛到单一模式——明明存在多条同样有效的解题路径，但模型只保留了一条。这使得 Agent 对分布外任务非常脆弱。

2. 梯度方差高 + 功劳归因不透明
多步骤任务中，哪一步的决策真正决定了最终成败？现有方法无法回答这个问题，信用分配只能靠模糊的整体奖励反向传播。

3. 技能进化缺乏原则性信号
”何时创建新技能、在哪创建、创建什么”——这三个问题现有方法只能靠启发式 LLM 判断，没有来自训练数据的数学支撑。

核心方法：三个流匹配信号

SkillFlow 的架构由三部分组成：可训练的 Supervisor Agent（负责动作选择）、冻结的 Executor（执行委托推理和工具调用）、动态技能库（根据训练诊断自主演化）。

回火轨迹平衡（Tempered Trajectory Balance，TTB）

这是 SkillFlow 的核心创新。传统流匹配让轨迹概率等于归一化奖励，TTB 改为让轨迹采样正比于奖励而非收敛到单一最优策略。

效果：保留了多条同等有效的解题路径，对抗了策略坍缩。TTB 同时联合训练前向策略（动作选择）和后向策略（基于事后条件的诊断）。

步骤重要性 I(t)

从流比率中推导出每一步的重要性分数，直接指向”哪个决策真正驱动了成功”。关键在于：后向策略提供这个信号的成本为零——不需要额外推理，信用归因免费附带。

技能边际流 F̂(s)

通过累积生成函数对技能的贡献排序，量化每个技能在任务成功中的边际价值。三个诊断信号合并回答：

• 何时进化（TTB 残差触发）
• 在哪进化（步骤重要性定位）
• 进化什么（技能边际流排名）

实验结果：14 个基准，全面超越

评测覆盖 7 个域内（IID）+ 7 个域外（OOD）基准，包括问答（HotpotQA）、数学推理（AIME）、代码生成（SWE-bench）、交互决策（WebShop、ALFWorld、ScienceWorld）。

主干模型为 Qwen3.5-9B，对比基线为 FlowSteer：

基准	SkillFlow	FlowSteer	提升
HotpotQA EM	92.19%	61.88%	+30.31%
WebShop 成功率	93.75%	55.94%	+37.81%
ALFWorld 成功率	96.09%	74.22%	+21.87%
SWE-bench Resolved	52.34%	39.06%	+13.28%
IID 平均 EM	94.14%	52.94%	+41.20%

同时，SkillFlow 实现了所有基线中最低的计算成本，token 用量和推理时间均下降 32-35%。准确性、多样性、效率三者同步优化，而非相互权衡。

消融实验验证了每个组件的独立价值

• 去掉 TTB：多样性敏感任务（AIME、WebShop、ALFWorld）退化最严重，直接证明了模式坍缩是真实失败原因
• 去掉后向策略：多步骤任务（AIME、WebShop、ScienceWorld）受损远大于事实类问答
• 去掉三个进化信号中的任意一个：都独立导致性能下降

反直觉的发现

OOD 优势反而更大：在分布外任务上，SkillFlow 对 REINFORCE 基线的领先幅度大约是 IID 的 1.5 倍。冻结的进化后技能似乎捕捉到了”可迁移的编排原语”，而不是特定基准的模板。

弱模型获益更多：基础能力较弱的 LLM 从显式信用归因和多样性保留采样中获益比例更高。这意味着 SkillFlow 对算力受限场景特别有价值——它在某种程度上弥补了基础推理能力的不足。

资源与使用

项目代码已在 GitHub 开源，支持 Python 3.10 + LoRA 高效微调，提供与 OpenAI 兼容的本地推理接口。训练数据（3500 条，跨 7 个基准家族）和 Qwen3.5-9B 监督检查点均可在 Hugging Face 获取。

作者团队：Mingda Zhang（港中深）、Tiesunlong Shen（新加坡国立大学）、Haoran Luo（南洋理工大学）、Wenjin Liu（南洋理工大学）、Zikai Xiao（浙江大学）、Erik Cambria（南洋理工大学）、Xiaoying Tang（港中深）

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

How should an agent’s skill library be managed? Letting an LLM decide heuristically when to create or prune skills often amounts to guesswork. SkillFlow provides a mathematically principled answer using flow-matching training signals — beating all baselines on 14 benchmarks with +41.2% average EM while simultaneously cutting compute costs by 32-35%.

📌 Paper: SkillFlow: Flow-Driven Recursive Skill Evolution for Agentic Orchestration
arXiv:2605.14089 full text: https://arxiv.org/abs/2605.14089
Code: https://github.com/beita6969/SkillFlow
Project: https://skill-flow.org

Three Fundamental Problems with Existing Approaches

1. Strategy Collapse Under Reward Maximization
Traditional RL training converges to a single mode. Multiple equally valid solution paths exist, but the model retains only one — making agents fragile on out-of-distribution tasks.

2. High Gradient Variance and Opaque Credit Assignment
In multi-step tasks, which decision actually determined success? Existing methods can’t answer this; credit assignment relies on blurry whole-trajectory reward backpropagation.

3. Unguided Skill Evolution
When to create a new skill, where to focus, what to add — current methods rely on heuristic LLM judgments with no grounding in training-derived signals.

Core Method: Three Flow-Matching Signals

SkillFlow’s architecture combines a trainable Supervisor Agent (action selection), a frozen Executor (delegated reasoning and tool calls), and a dynamic skill library that evolves based on training diagnostics.

Tempered Trajectory Balance (TTB)

TTB is the central innovation. Standard flow matching sets trajectory probability equal to normalized reward, converging to a single optimum. TTB instead samples trajectories proportional to reward, preserving multiple equally effective solution paths and countering strategy collapse.

TTB jointly trains a forward policy (action selection) and a backward policy (hindsight-conditioned diagnostics).

Step Importance I(t)

Derived from flow ratios, this score identifies which specific decisions drove success. The key: the backward policy provides this signal at zero additional inference cost — transparent credit attribution comes for free.

Skill Marginal Flow F̂(s)

Ranks each skill’s contribution to task success via cumulant generating functions. The three diagnostics together answer:

• When to evolve (TTB residual floor triggers)
• Where to focus (step importance signals)
• What to add (skill marginal flow rankings)

Results: 14 Benchmarks, Across the Board

Evaluation covers 7 in-distribution (IID) + 7 out-of-distribution (OOD) benchmarks including HotpotQA, AIME, SWE-bench, WebShop, ALFWorld, and ScienceWorld. Backbone: Qwen3.5-9B. Baseline: FlowSteer.

Benchmark	SkillFlow	FlowSteer	Gain
HotpotQA EM	92.19%	61.88%	+30.31%
WebShop Success	93.75%	55.94%	+37.81%
ALFWorld Success	96.09%	74.22%	+21.87%
SWE-bench Resolved	52.34%	39.06%	+13.28%
Average IID EM	94.14%	52.94%	+41.20%

SkillFlow also achieves the lowest compute cost of all baselines — 32-35% fewer tokens and less inference time. Accuracy, diversity, and efficiency improve simultaneously rather than trading off against each other.

Ablations Validate Each Component

• Removing TTB: Largest degradation on diversity-sensitive tasks (AIME, WebShop, ALFWorld) — confirms mode collapse as the real failure mode
• Removing the backward policy: Multi-step tasks (AIME, WebShop, ScienceWorld) suffer more than factual QA
• Removing any of the three evolution signals: Each independently degrades performance

Counterintuitive Findings

OOD advantage widens: SkillFlow’s lead over REINFORCE baselines is roughly 1.5× larger on out-of-distribution tasks than on IID. The frozen evolved skills appear to capture transferable orchestration primitives rather than benchmark-specific templates.

Weaker models gain more: LLMs with weaker base reasoning benefit disproportionately from explicit credit attribution and diversity-preserving sampling — making SkillFlow particularly valuable for compute-constrained deployments.

Resources

Code is open-sourced on GitHub, supporting Python 3.10 with LoRA fine-tuning and OpenAI-compatible local inference. Training data (3,500 records across 7 benchmark families) and the Qwen3.5-9B supervisor checkpoint are on Hugging Face.

Authors: Mingda Zhang (CUHK-Shenzhen), Tiesunlong Shen (NUS), Haoran Luo (NTU), Wenjin Liu (NTU), Zikai Xiao (Zhejiang University), Erik Cambria (NTU), Xiaoying Tang (CUHK-Shenzhen)

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.